The NLL-vs-ECE argument is the strongest part: ECE jumping at bin edges makes it a bad search target, and the q = 0.92 table shows why NLL lands on the true rate. One thing I'd be curious about is the step-wise fit on only 729 right-group answers for the intent temperature. Since the intent step never sees wrong-group cases at fit time but is scaled on them at test time, would a held-out check on wrong-group answers alone show whether T=1.14 over- or under-corrects there? Also, since the middle bins stay >0.1 off, a two-parameter option like Platt/vector scaling or isotonic regression on the final product might be worth comparing, with a bootstrap interval on the test ECE to see whether 0.041→0.030 is outside the noise.